Path-Selection Heuristics for Dominator-Path Scheduling

Virtually all programs exhibit some form of parallelism. Instruction-level parallel (ILP) architectures are capable of exploiting the fine-grained parallelism inherent in most programs by simultaneously executing multiple operations from the same instruction stream. This type of architecture requires special compilation and optimization techniques; one such optimization, instruction scheduling, reorders the instructions in a function in an effort to overlap their latency and thereby speed up the entire function. Dominator-Path Scheduling (DPS), a popular scheduling method, moves code along a selected path in the dominator tree. There are many ways to select these paths; this thesis presents an investigation of the characteristics of good choices of dominator paths for use in DPS. No single heuristic performed well enough to use on its own, but a combined heuristic can do very well in almost all cases. I recommend a combination of three heuristics: limiting paths to one nesting level of a single loop, grouping sparse blocks together, and including non-postdominating blocks. This combined heuristic achieved the best possible schedules in the benchmarks bubble and matrix_mult and within 4% of best in gauss. For bubble, this heuristic was 24% better than DPS’s previous method of choosing dominator paths. The combined heuristic is likely to do as well as is possible in most functions.